So, there's a lot of different points I want to get to, but I want to start with this blog post that you put out recently, and it talked about your affinity for open-source models. And I was struck by this line. You wrote that you're using open-source AI models at Pinterest to achieve similar performance at less than 10% of the cost of leading proprietary AI models. Look, we know how competitive these models are. That's a big claim. What did you mean by that? Talk about that.
Yeah, I think this is, you know, so much of the discussion around AI has been about who's building the latest largest proprietary model. And I think there's a couple other really important trends that are happening that aren't being talked about enough. The one is open source. And as we published, we're seeing that we can take open-source models and fine-tune those and get similar performance to the very best proprietary models at less than 10% of the cost. And I think that's really really important because open source has a democratizing effect that has been true for the building of software for decades now. And I think that's really important for how we look forward as to how many many companies can build with AI. If you can get similar capabilities for less than 10% of the cost, that's going to have a really democratizing effect, really spur a lot of innovation, creativity. And so we're pushing our work on that to share with others just how good that is because we think it's really important there's a thriving open-source community for AI. And then there's a second trend around compact fit-for-purpose models that I think also is about you don't need the very largest models in the world to do every single task. And that's also a way that you can create cost effectiveness in AI as well.
So I just want to make sure that I understand this. I mean somebody might hear this and think, oh, less than 10% of the cost of leading AI models. I mean we think about Google and OpenAI as the leaders with their proprietary models. So help me understand. Are you saying that the models that Pinterest has built using open source technology is just as good and cheaper? Is that what you're saying?
Two things. One, there are large scale open-source models available now that you can take and fine-tune them for your purpose. And what we found for our purposes is that we could achieve 90% lower cost at similar performance on the things that we needed to do.
Okay. And so for a very specific use case.
Well, yes. And but I think this is a generalizable point because that's about the large generalized models. There's a second point which is related to this that we are also building our own in-house models for specific tasks. And so these are two different tactics that can help companies get a lot more out of AI at much lower cost. And this is really important right now because nearly every CEO that I talked to, for example, I was last week at the New York Times DealBook and so many of the CEOs that I talked to there said, well, they've invested a lot of money in AI and buying these off-the-shelf proprietary software solutions, but they're not seeing the return on investment that they need. They're spending a lot but they're not yet getting the savings that they had expected or the spend is significantly offsetting the savings that they were getting.
And companies finding that they can leverage open source to get much better efficiency. This has been true in software development for decades. In fact, many of the largest companies out there wouldn't exist if not for the fact that they have been able to build on open source. It was much cheaper than proprietary databases or proprietary operating systems. Things like Linux and those types of things were really, really important to the building of a lot of these trillion dollar market cap companies. Well, the next batch of companies, they're going to need to have really great open-source AI available to them to have these cost efficiencies. Otherwise, the proprietary software may collect all the value without others being able to go build and create value from that as well. So we think this is really important again which is why we're sharing our findings.
So you're a big proponent of open source. What do you think of the open source models coming out of China? I mean that's in the news this week. Tons of progress there earlier in the year. We've sort of heard less about them but we know that it's just one model away from rocking markets as we've seen. Are you sort of how closely are you watching that space?
Yeah, I mean the DeepSeek moment was a huge moment and I think so much of the DeepSeek moment the conversation was about how they had done that on less than the latest GPU chips and the sort of theory of constraints of how they leveraged those constraints to find cheaper ways to build those models. And that was really about the chips but there were a lot of techniques they used that are now going into the open source community that others can use to build their own models. So we're using some of those techniques for our own models that we're building internally. But then you've had other really good open source models emerge like Qwen for example that is putting out really powerful models. But I think that DeepSeek moment was so much of the focus was on the chips and lower cost chips. But I think the bigger moment was that you had an open-source model that was rivaling the large proprietary models. And now you have that with Qwen. You see others coming out, Mistral, like you're seeing an open source community really start to develop that can compete. Open source is at the table and it is competing and doing so at really effective cost to performance levels. And that's important not just from a cost perspective. I think it's also important that so many people refer to these models as this model came from this country or that model came from that country. The really important thing about open source is that open source belongs to everyone. It's not controlled by anyone company. True open source is not controlled by anyone company. It's not controlled by any one nation state. And if you want open AI to be in the hands of the many rather than the few, then open source is really important to that. And again that has been really important to software development for decades now. And I think it's really important that that continues in the world of AI, right? And it doesn't need to be instead of proprietary models. Open source software has been huge for decades. It doesn't mean that proprietary software no longer exists. It just means that open source software is a really important component of a thriving ecosystem overall alongside of proprietary models and can compete well with those proprietary models. It's a good check and balance on those proprietary models.
So let me ask you a question broadly about Pinterest and how you're approaching AI now. Where are you applying the gas on AI and then where are you pushing the brakes? Because we've done some reporting at The Information and we've spoken to it was your CTO actually who spoke with us about ways in which the company is taking a more careful approach to for example AI generated content. Talk to me a little bit about how you see that.
Well, we're applying the gas. We've effectively turned Pinterest into an AI powered shopping assistant over the last few years. We've had nine straight quarters of record high users. More than half the platform is Gen Z. Shopping is the primary reason they come to the platform. And if you ask Gen Z users why they come to Pinterest, they'll say things like, 'Well, Pinterest just gets me.' And that's us using our AI models tuned off our unique curation signal to give really great personalized recommendations and agentic style experiences that help guide users through shopping journeys. So we're really doubling down on that. We've just launched our latest Pinterest assistant going even further with that. But we're really building shopping experiences that don't replace shopping or automate away shopping. Right. I see folks that are doing that, trying to build shopping for people that hate shopping. We're building shopping assistance for people that love shopping. And they want an assistant to actually help them immerse in that journey. So that's where we're doubling down and hitting the gas. To your point, the place where I wouldn't say we're hitting the brakes. I'd say we're just making sure that we use AI responsibly. The first is tuning AI for positivity. This is one of the primary reasons I joined Pinterest from Google nearly three and a half years ago is that I wanted to prove there was a more positive alternative to the business model of social media that so much of it had engagement via engagement at the core. And a big part of that, how did social media become negative? AI got put in charge of what you see on social media more than a decade ago. It was just earlier forms of AI. And this is like avoiding slop essentially. Well, avoiding slop or avoiding other negative things like when AI was asked to maximize your view time on social media more than a decade ago, the AI figured out you look longer at things that trigger you. Whatever your triggers are. And so we set out to tune AI for positivity. If AI could be used to keep you glued to a screen with engagement via enragement, why can't we ask the AI to make sure you leave the platform feeling better and it's a positive contributor to your emotional well-being? And we've been able to prove that out. So, that's a place where we're making sure we don't have a race to the bottom on just trying to keep people glued to a screen doing things that are addictive rather than additive. And so, we try to focus on additive. That's one place. Another place is on trust and safety where we're using AI to combat bad content. And you asked about AI slop, well there's been a huge increase in the ability to create content and the vast majority of that are people that are having fun creating good content. You've had a democratization of people's ability to express themselves and that's a good thing. But also mixed in that like any technology, generative AI can be used for good or for bad. And so you also have bad actors that are trying to create things that are spammy or not helpful or harmful.
So is there AI generated content right now on Pinterest that is allowed?
Oh, absolutely. And what I would say with AI generated content, this whole discussion of AI slop, a couple things I'd say. One is that it's not about is AI content good or bad. It's about some AI content is good and some AI content is bad. And how do you parse that? And how do you give the user control over what content they want to see?
